Metal powder particle size statistical method and system based on artificial intelligence

Abstract

This invention relates to the field of particle image recognition technology, specifically to an artificial intelligence-based method and system for statistical analysis of metal powder particle size. The method includes sample preprocessing and image acquisition to obtain a high-resolution image containing scale calibration marks; preprocessing the high-resolution image with scale calibration marks to obtain a preprocessed grayscale image; using a dual-branch architecture HDP-DSL algorithm to identify and segment the preprocessed grayscale image, obtaining a full-scale high-precision particle segmentation mask; calculating single-particle physical feature parameters based on the full-scale high-precision particle segmentation mask; performing statistical analysis on the single-particle physical feature parameters to generate a particle size distribution histogram and calculate key statistical feature values ​​such as D10, D50, and D90; and visualizing the detection results through color coding and interactive design, thus solving the problems of difficult particle segmentation due to overlapping particles and low recognition rate due to blurred edges when identifying high-density metal powder.

Description

Technical Field

[0001] This invention relates to the field of particle image recognition technology, specifically to an artificial intelligence-based method and system for statistical analysis of metal powder particle size. Background Technology

[0002] Metal powders are important raw materials in additive manufacturing (such as laser powder bed melting), powder metallurgy and other fields. The particle size distribution of metal powders directly affects the flowability, bulk density, sintering behavior of the powders, as well as the mechanical properties and surface quality of the final products. Accurate determination of particle size distribution is crucial for understanding and controlling the performance of metal powders.

[0003] Currently, mainstream methods for testing the particle size distribution of metal powders include sieving, laser diffraction, and image analysis. Sieving is simple to operate and low in cost, but it has low accuracy for fine particles smaller than 45 micrometers and is time-consuming. Laser diffraction is fast and has a wide measurement range, but it is easily affected by particle shape and tends to overestimate the proportion of ultrafine particles for irregular particles. It also requires strict control of sample dispersion conditions. Image analysis can directly obtain particle morphology and size information and is the most intuitive particle size characterization method, making it the preferred method for refined testing in the industry.

[0004] The granularity statistics of existing image analysis methods mainly rely on commercial software such as ImageJ. These software programs use traditional image segmentation and morphological measurement methods, and achieve particle recognition by manually setting thresholds and outlining contours. This requires high image quality, and threshold adjustment requires multiple iterative operations, making the process cumbersome and inefficient. At the same time, traditional software only supports single-image analysis and cannot achieve automatic batch processing of multiple images. A large number of images need to be analyzed to obtain statistically significant results, resulting in high manual costs and long detection cycles.

[0005] However, existing image analysis methods have significant drawbacks when processing high-density metal powders. The particles overlap severely, and at high magnification, the particles are often tightly packed or even overlapping each other. Traditional threshold segmentation methods are unable to accurately distinguish the boundaries of adjacent particles, resulting in large statistical deviations in particle size. At the same time, in high-density particle scenes, the distance between particles is extremely small, and the boundaries are blurred. Traditional single threshold segmentation methods are unable to accurately distinguish between particles and the background. Summary of the Invention

[0006] This invention provides an artificial intelligence-based method and system for statistical analysis of metal powder particle size, aiming to solve the problems of difficult segmentation of overlapping particles and low recognition rate of blurred edges when identifying high-density metal powder.

[0007] The first aspect of this invention aims to provide an artificial intelligence-based method for statistical analysis of metal powder particle size, comprising the following steps:

[0008] Step 1: Sample preprocessing and image acquisition to obtain high-resolution images containing scale calibration marks;

[0009] Step 2: Preprocess the high-resolution image containing scale calibration marks to obtain a preprocessed grayscale image;

[0010] Step 3: The HDP-DSL algorithm with a dual-branch architecture is used to identify and segment the preprocessed grayscale image to obtain a full-scale high-precision particle segmentation mask.

[0011] Step 4: Calculate the physical feature parameters of a single particle based on the full-scale high-precision particle segmentation mask;

[0012] Step 5: Perform statistical analysis on the single-particle physical characteristic parameters output in Step 4, generate a particle size distribution histogram, and calculate key statistical characteristic values ​​such as D10, D50, and D90.

[0013] Step 6: Visualize the detection results through color coding and interactive design.

[0014] Furthermore, step 1 specifically includes:

[0015] Anhydrous ethanol or acetone is used as a solvent, and the sample is ultrasonically cleaned for 5-10 minutes with a power of 300-500W. After cleaning, the sample is placed in a vacuum drying oven and dried for 2-4 hours at 40-60℃ and a vacuum degree of <10Pa to ensure that the solvent is completely evaporated and there is no residue.

[0016] Take 1-5 mg of dried metal powder and disperse it evenly on an aluminum or copper conductive adhesive sample stage. If the powder particle size is less than 1 μm, it can be dispersed in anhydrous ethanol to prepare a suspension. Use a micropipette to take 1-2 μL of the suspension and drop it onto the conductive adhesive on the sample stage and vacuum dry it.

[0017] The powder samples were photographed and observed using a scanning electron microscope. The overall dispersion of the samples was observed at a low magnification of 50-200x, avoiding areas with severe agglomeration or sparse distribution. The magnification was then gradually increased to 500-5000x, and 10-20 representative fields of view were selected for photographing. All photographs were labeled with scale bars. The image resolution was set to 1024×1024 or 2048×2048 pixels and saved in TIFF or PNG format to obtain high-resolution images with scale calibration marks.

[0018] Furthermore, step 2 specifically includes:

[0019] First, the high-resolution image containing scale calibration marks is converted into a grayscale image using an RGB channel weighted fusion algorithm, and then Gaussian blur is applied to the grayscale image to suppress noise.

[0020] The Laplacian operator is used to enhance the edges of the denoised image. The mathematical expression for edge enhancement is as follows:

[0021]

[0022]

[0023] The edge-enhanced image is combined with the original image through weighted fusion to obtain a preprocessed grayscale image.

[0024] Furthermore, in step 3, the HDP-DSL algorithm consists of two core modules: a large-size particle detection and segmentation mechanism based on geometric features, and a small-size edge particle detection and segmentation algorithm based on an improved U-Net neural network.

[0025] The specific operations of the large-size particle detection and segmentation mechanism based on geometric features include:

[0026] The preprocessed grayscale image is binarized using both global and adaptive local thresholding, and the union of the two thresholding results is taken to obtain candidate contours.

[0027] For candidate contours, a distance transformation map is obtained using the distance transformation formula.

[0028] The formula for calculating the distance transformation is:

[0029]

[0030]

[0031] The local maximum detection algorithm is used to find peak points in the distance transformation map. The minimum peak detection distance is set to dmin=0.3max(D(x,y)); if the number of valid peaks detected is greater than 1, the contour is determined to be overlapping particles.

[0032] For the identified overlapping particles, the overlapping region is segmented into independent particle regions by watershed transformation to obtain the segmentation result of the first branch;

[0033] Small-size and edge particle detection and segmentation based on an improved U-Net neural network:

[0034] The preprocessed grayscale image is input into the improved U-Net neural network to obtain the segmentation result of the second branch;

[0035] The segmentation results of the first branch and the segmentation results of the second branch are complementary and corrected to obtain a full-scale high-precision particle segmentation mask.

[0036] Furthermore, the improved U-Net neural network is a neural network optimized based on the classic U-Net network architecture;

[0037] The encoder in the classic U-Net network architecture is optimized by replacing the standard convolution of the classic encoder with depthwise separable convolution.

[0038] After the encoder feature output, a spatial attention module and a channel attention module are added;

[0039] Dice Loss and binary cross-entropy loss function are used.

[0040] Furthermore, step 4 specifically includes:

[0041] For each segmented effective particle profile, calculate characteristic parameters such as area, equivalent diameter, caliper diameter, and center coordinates.

[0042] The equivalent diameter is calculated based on the principle of area equivalence, and the formula is: ;

[0043] Where deq is the equivalent diameter and A is the particle area.

[0044] The diameter of the caliper is obtained using the minimum circumscribed rectangle algorithm; the center coordinates are obtained using the minimum circumscribed circle algorithm.

[0045] Calibration is performed using scale calibration markers in the image, converting pixel-level parameters to physical units.

[0046] Furthermore, step 5 specifically includes:

[0047] Using particle caliper diameter as the core parameter, calculate the minimum, maximum, average, and standard deviation of particle size;

[0048] The statistical characteristic values ​​of D10, D50, and D90 were calculated using the quantile statistical method.

[0049] The particle size range is divided into multiple intervals, the number of particles in each interval is counted, and a particle size distribution histogram is generated.

[0050] A second aspect of this invention aims to provide an artificial intelligence-based metal powder particle size statistics system, comprising:

[0051] The image acquisition module is configured to acquire preprocessed sample images and obtain high-resolution images containing scale calibration marks;

[0052] The image preprocessing module is configured to preprocess a high-resolution image containing scale calibration marks to obtain a preprocessed grayscale image;

[0053] The particle detection and segmentation module is configured to use the HDP-DSL algorithm with a dual-branch architecture to identify and segment the preprocessed grayscale image to obtain a full-scale high-precision particle segmentation mask.

[0054] The parameter calculation module is configured to calculate the physical feature parameters of a single particle based on a full-scale high-precision particle segmentation mask.

[0055] The statistical analysis module is configured to perform statistical analysis on the physical characteristic parameters of single particles, generate a particle size distribution histogram, and calculate key statistical characteristic values ​​of D10, D50, and D90.

[0056] The visualization and interaction module is configured to visualize the detection results through color coding and interactive design.

[0057] The beneficial effects achieved by this invention are as follows: Image preprocessing, through a combination of grayscale conversion, Gaussian filtering, dual-threshold fusion, and edge enhancement, combined with core mathematical formulas, achieves precise processing, effectively removing noise and enhancing contrast, laying a high-quality data foundation for subsequent recognition and segmentation; Particle recognition introduces an improved U-Net architecture, which enhances key information through automatic extraction of high-order features and attention mechanisms, significantly improving the accuracy of distinguishing particles from the background, with strong generalization ability, adapting to the detection needs of different types of powder particles; Parameter calculation simultaneously obtains multi-dimensional parameters such as area, equivalent diameter, and caliper diameter, combined with high-precision scale calibration, comprehensively reflecting particle morphological characteristics; Statistical analysis generates basic statistical indicators, quantile feature values, and particle size distribution histograms, meeting the analysis needs of different scenarios and providing a reliable basis for quality assessment; Interactive visualization intuitively presents particle size distribution through color coding, bidirectional linkage positioning facilitates key observation, and batch processing of multiple images improves analysis efficiency and reduces labor costs. Attached Figure Description

[0058] Figure 1 This is a flowchart of the artificial intelligence-based metal powder particle size statistical method of the present invention.

[0059] Figure 2 This is a schematic diagram of the artificial intelligence-based metal powder particle size statistics system of the present invention.

[0060] Figure 3 This is a schematic diagram of the particle segmentation effect of the present invention.

[0061] Figure 4 This is a schematic diagram of the improved U-Net neural network structure of the present invention.

[0062] Figure 5 This is a schematic diagram of the result visualization interface of the present invention. Detailed Implementation

[0063] To facilitate understanding of the present invention by those skilled in the art, specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0064] like Figure 1-2 As shown, an artificial intelligence-based method for statistical analysis of metal powder particle size includes the following steps:

[0065] Step 1: Sample preprocessing and image acquisition to obtain high-resolution images containing scale calibration marks;

[0066] The specific operations for sample preprocessing and image acquisition include:

[0067] First, the metal powder sample is pretreated to remove surface impurities and residual solvents. Anhydrous ethanol or acetone is used as the solvent, and the sample is ultrasonically cleaned at 300-500W for 5-10 minutes to effectively remove oil, oxide layer and impurity particles from the sample surface. After cleaning, the sample is placed in a vacuum drying oven and dried at 40-60℃ and a vacuum degree <10Pa for 2-4 hours to ensure that the solvent evaporates completely without residue.

[0068] Next, the dried sample is prepared to ensure its stability for subsequent observation. Take 1-5 mg of dried metal powder and disperse it evenly on an aluminum or copper conductive adhesive sample stage. Gently press to ensure close contact between the powder and the conductive adhesive to reduce sample drift during electron beam irradiation. If the powder particle size is less than 1 μm, it can be dispersed in anhydrous ethanol to prepare a suspension. Use a micropipette to take 1-2 μL of the suspension and drop it onto the conductive adhesive on the sample stage. After vacuum drying, proceed with the subsequent steps to achieve uniform dispersion of the powder.

[0069] Finally, the powder samples were photographed and observed using a scanning electron microscope. First, the sample stage was moved to observe the overall dispersion of the sample at a low magnification of 50-200x, avoiding areas with severe agglomeration or sparse distribution. Then, the magnification was gradually increased to 500-5000x, and 10-20 representative fields of view were selected for photography. All photos were labeled with scale bars. The image resolution was set to 1024×1024 or 2048×2048 pixels to meet the needs of subsequent particle size statistics and other analyses. High-resolution images (TIFF or PNG) with scale calibration marks were output.

[0070] Step 2: Preprocess the high-resolution image containing scale calibration marks to obtain a preprocessed grayscale image;

[0071] In high-density particle scenes, the distance between particles is extremely small and the boundaries are blurred, making it difficult for traditional single threshold segmentation methods to accurately distinguish particles from the background.

[0072] First, the high-resolution image containing scale calibration marks is converted into a grayscale image using an RGB channel weighted fusion algorithm, and then Gaussian blur is applied to the grayscale image to suppress noise.

[0073] The Laplacian operator is used to enhance the edges of the denoised image, thereby improving particle boundary information. The mathematical expression for edge enhancement is as follows:

[0074]

[0075]

[0076] The Laplacian operator can detect the second derivative of an image and is highly sensitive to regions with significant intensity changes, such as particle boundaries.

[0077] The edge-enhanced image is combined with the original image through weighted fusion to obtain a preprocessed grayscale image; this preserves the details of the original image while highlighting edge features, thus solving the problem of blurred boundaries in high-density grain scenes.

[0078] Step 3: The HDP-DSL algorithm with a dual-branch architecture is used to identify and segment the preprocessed grayscale image to obtain a full-scale high-precision particle segmentation mask.

[0079] This invention proposes a High-Density Particle Detection and Segmentation Learning Algorithm (HDP-DSL) for high-density metal powder particle scenes. This algorithm combines a geometric feature-based overlapping particle detection mechanism with an improved U-Net neural network architecture, enabling accurate identification and segmentation of high-density overlapping particles.

[0080] The algorithm mainly consists of two core modules: a large-size particle detection and segmentation mechanism based on geometric features (first branch) and a small-size edge particle detection and segmentation algorithm based on an improved U-Net neural network (second branch). These two modules work together to solve the problems of blurred edges and difficulty in segmenting overlapping particles in high-density scenes, as well as model overfitting under small sample conditions in traditional methods.

[0081] The specific operations of the large-size particle detection and segmentation mechanism based on geometric features include:

[0082] The preprocessed grayscale image is binarized using both global thresholding and adaptive local thresholding. The union of the two thresholding results is then taken to achieve comprehensive detection of large particles. Global thresholding is suitable for particles with high contrast in the central region, while adaptive local thresholding can adapt to the gradual changes in the edge region. The final binarization result is the union of the two thresholding methods to ensure that particles under various lighting conditions in high-density scenes can be detected correctly.

[0083] For the identified candidate contours, a distance transformation map is obtained using the distance transformation formula.

[0084] The formula for calculating the distance transformation is:

[0085]

[0086]

[0087] The local maximum detection algorithm is used to find peak points in the distance transformation map. The minimum peak detection distance is set to dmin=0.3max(D(x,y)). This can avoid false peaks caused by noise and accurately identify the center point of overlapping particles. If the number of valid peaks detected is greater than 1, the contour is determined to be an overlapping particle.

[0088] The distance transformation inside the contour is calculated using the distance transformation formula. The distance transformation can reflect the shortest distance from each pixel to the contour boundary. For a single circular or nearly circular particle, its distance transformation map will show a single-peak distribution, with the peak located at the center of the particle. For multiple overlapping particles, the distance transformation map will show a multi-peak distribution, with each peak corresponding to a potential particle center.

[0089] For the identified overlapping particles, this invention uses an improved watershed algorithm for segmentation. The local maximum point of the distance transformation is used as the marker point. The overlapping area is segmented into independent particle regions through watershed transformation to obtain the segmentation result of the first branch. In order to handle ultra-large adhesion areas (size exceeding 2000 pixels), the algorithm will automatically perform intelligent scaling of the region to maintain segmentation accuracy while ensuring computational efficiency.

[0090] Compared to traditional watershed algorithms, distance transformation can calculate the distance from each foreground pixel to the nearest background pixel, and its local maximum point corresponds exactly to the core region of each particle. Traditional watershed algorithms often use simple thresholding and morphological operations (such as erosion) to obtain marker points. These methods are easily affected by noise, resulting in either too many marker points (leading to oversegmentation) or too few marker points (leading to undersegmentation).

[0091] Small-size and edge particle detection and segmentation based on an improved U-Net neural network:

[0092] The preprocessed grayscale image is input into the improved U-Net neural network to obtain the segmentation result of the second branch;

[0093] The segmentation results from the first branch and the second branch are complementary and corrected to obtain a full-scale, high-precision particle segmentation mask. The segmentation effect is as follows: Figure 3 As shown, particles of different sizes can be clearly identified, with no missed or false detections.

[0094] The segmentation results of the first branch and the second branch suffer from inconsistent accuracy and complementary recognition ranges. By clarifying the fusion judgment criteria, conflict resolution rules, and result integration process, complementary correction of the results from the two segmentation modules is achieved, ultimately outputting a full-scale, high-precision granular segmentation mask. The specific implementation is as follows:

[0095] Integration prerequisites and judgment criteria (clarifying the scope of integration and conflict scenarios)

[0096] First, particle matching is determined and valid results are screened.

[0097] Particle matching determination: The Intersection over Union (IOU) of pixel coordinates is calculated to determine whether the recognition results of the two branches correspond to the same particle; the IOU threshold is set to 0.5; when the IOU of the large particle contour recognized by the first branch and the same particle contour recognized by the second branch is ≥0.5, it is determined to be "the same particle", which belongs to the fusion conflict scenario; when the IOU is <0.5, it is determined to be "different particles", which belongs to the result complementary scenario.

[0098] Valid result determination: To ensure fusion quality, invalid results are removed in advance; for large particles identified by the first branch, both "fill rate ≥ 0.7" and "contour area ≥ 50 pixels" must be met to exclude noise false detection; for particles identified by the second branch, "prediction probability ≥ 0.8" must be met to exclude pseudo-segmentation regions; recognition results that do not meet the above conditions are directly removed and do not participate in subsequent fusion.

[0099] Secondly, the following differentiated processing logic is executed for different scenarios:

[0100] Scenario 1: Dual-module recognition of the same particle (IOU≥0.5); In response to the problem that the segmentation results of the first branch have rough contours and incomplete edges, while the segmentation results of the second branch have fine contours but have small-scale pseudo-segments, a "basic range limitation + edge detail correction" strategy is adopted; Specifically: the particle contour recognized by the first branch is used as the "basic range", and pseudo-segmentation areas in the second branch recognition results that exceed the basic range are removed; then, the particle edge details recognized by the second branch segmentation results are used to correct the rough contours and incomplete edges in the segmentation results of the first branch.

[0101] Scenario 2: The first branch misses small particles, while only the second branch identifies the particle;

[0102] For small particles, verify whether the particle identified by the second branch meets the condition of "prediction probability ≥ 0.8"; if it does, directly include it in the final segmentation result to supplement the missed detection area of ​​the first branch; if it does not meet the condition, remove it directly to avoid introducing false segmentation.

[0103] Scenario 3: The second branch is missed, and only the first branch is identified;

[0104] For ultra-large particles, verify whether the particle identified by the first branch meets the requirements of "fill rate ≥ 0.7" and "outline area ≥ 50 pixels"; if it meets the requirements, directly include it in the final segmentation result to supplement the missed areas of module 2.

[0105] Scenario 4: Neither module was recognized;

[0106] These are considered unrecognizable particles or background noise and discarded directly to avoid misjudgments caused by forced recognition and to ensure the accuracy of the final segmentation results.

[0107] Final fusion output;

[0108] By integrating the processing results of the above four scenarios, a high-density particle segmentation result containing full-scale and high-precision information is obtained, realizing the complementary synergy between the first branch in controlling the overall shape of large particles and the second branch in recognizing small particles and edge details.

[0109] like Figure 4 As shown, the improved U-Net neural network is a neural network optimized based on the classic U-Net network architecture;

[0110] The classic U-Net network architecture uses an encoder-decoder structure. To adapt to small sample scenarios, this invention optimizes the encoder in the original U-Net network architecture by replacing the standard convolution of the classic encoder with depthwise separable convolution. This addresses the problem of large number of parameters in the standard convolution and the tendency to overfit in small sample training, thus improving the stability of segmentation accuracy. The depthwise separable convolution splits the standard convolution into two steps: "depthwise convolution (channel-wise convolution) + point convolution (channel fusion)". Without sacrificing feature extraction capabilities, it significantly reduces the number of model parameters, greatly reduces the risk of overfitting under small sample conditions, and improves model training efficiency, enabling it to quickly adapt to high-density particle segmentation tasks and output stable segmentation results. The improved encoder retains the downsampling structure (MaxPooling) of the classic U-Net and does not change its feature extraction process, only optimizing the convolution method.

[0111] After the encoder's feature output, a dual attention mechanism (spatial attention module + channel attention module) is added. This mechanism addresses the issues of weak feature selection ability and high background noise interference, reducing pseudo-segmentation. Spatial attention can accurately focus on the region where small particles are located and the details of particle edges, ignoring irrelevant background noise and improving the feature attention of subtle regions. Channel attention adaptively adjusts the weights of different feature channels, highlighting feature channels related to small particles and particle edges, while weakening background feature channels unrelated to particles. This significantly improves the model's feature representation ability with a small number of parameters, making it particularly suitable for small-sample learning scenarios and perfectly adapting to the segmentation needs of small particles and edge details.

[0112] The computational process of the attention mechanism is as follows: The spatial attention module first aggregates features along the channel dimension. :

[0113]

[0114]

[0115]

[0116] This dual attention mechanism can significantly improve the model's feature representation ability with a small number of parameters, making it particularly suitable for few-shot learning scenarios.

[0117] In terms of loss function design, this invention adopts a combination of Dice Loss and binary cross-entropy loss, which considers both region overlap and pixel-level classification accuracy. Furthermore, for the difficult-to-classify region of particle boundaries, a Focal Loss mechanism is introduced to give higher weight to difficult-to-classify samples. The final combined loss function is defined as follows:

[0118]

[0119] This multi-loss combination can effectively balance region overlap and boundary accuracy, achieving better training results under small sample conditions.

[0120] Step 4: Calculate the physical feature parameters of a single particle based on the full-scale high-precision particle segmentation mask;

[0121] For each segmented effective particle profile, calculate characteristic parameters such as area, equivalent diameter, caliper diameter, and center coordinates.

[0122] The equivalent diameter is calculated based on the principle of area equivalence, and the formula is: ;

[0123] Where deq is the equivalent diameter and A is the particle area; this facilitates comparison with traditional particle size test results.

[0124] The caliper diameter is obtained by using the minimum bounding rectangle algorithm to find the minimum bounding rectangle of the particle. The length of the longer side of the rectangle is taken as the caliper diameter, reflecting the maximum size characteristics of the particle. The center coordinates are obtained by using the minimum bounding circle algorithm, providing a basis for subsequent visualization and positioning.

[0125] To convert pixel-level parameters to physical units (micrometers, μm), calibration is performed using scale calibration markers (micrometer-level scales of known size) in the image. First, the scale region is located in the image, and the length pixel value of the scale is extracted through threshold segmentation and contour detection. Then, the pixel-to-physical unit conversion coefficient is calculated based on the actual length of the scale, achieving a calibration accuracy of 0.001 μm / pixel. Finally, all feature parameters are multiplied by the conversion coefficient to obtain physical parameters in micrometers, ensuring the practicality of the measurement results.

[0126] To improve the accuracy of parameter calculation, a parameter verification mechanism is set up to verify the consistency between the calculated equivalent diameter and the caliper diameter. If the difference between the two exceeds 30%, the contour segmentation results are re-examined to eliminate parameter anomalies caused by segmentation errors. At the same time, the rationality of extreme values ​​with excessively large or small areas is judged, and the actual characteristics of the powder material are combined to eliminate obviously abnormal parameter values ​​to ensure the reliability of the final output parameters.

[0127] Step 5: Perform statistical analysis on the single-particle physical characteristic parameters output in Step 4, generate a particle size distribution histogram, and calculate key statistical characteristic values ​​such as D10, D50, and D90.

[0128] Statistical analysis is a key step in comprehensively evaluating a large number of particle parameters. This invention comprehensively reflects the characteristics of powder particle size distribution through multi-dimensional statistics. First, the characteristic parameters of all effective particles (mainly caliper diameter) are statistically analyzed to calculate basic indicators such as minimum, maximum, average, and standard deviation. The average value reflects the average particle size, and the standard deviation reflects the dispersion of particle size distribution, providing a basis for evaluating material uniformity.

[0129] Key characteristic values ​​such as D10, D50, and D90 are calculated using quantile statistics: D10 represents the particle size at which the cumulative particle size distribution reaches 10%, reflecting the distribution characteristics of fine particles; D50 (median diameter) represents the particle size at which the cumulative particle size distribution reaches 50%, and is the core indicator characterizing the average particle size; D90 represents the particle size at which the cumulative particle size distribution reaches 90%, reflecting the distribution characteristics of coarse particles. These characteristic values ​​are important bases for the quality control of powder materials and meet industry standards and practical application requirements.

[0130] Based on the distribution range of particle caliper diameter, the number of bins in the histogram is adaptively adjusted (usually 30 groups) to divide the particle size range into multiple intervals. The number of particles in each interval is counted to generate a particle size distribution histogram. The horizontal axis of the histogram represents the particle size interval (μm), and the vertical axis represents the number of particles, intuitively presenting the proportion of particles in different particle size intervals, providing support for users to quickly grasp the particle size distribution pattern.

[0131] Step 6: Visualize the detection results through color coding and interactive design.

[0132] Visualizing the results is key to improving system usability. This invention employs color coding and interactive design to achieve intuitive display and flexible operation of particle distribution. Based on the caliper diameter of the particles, a gradient color mapping mechanism is used for color coding, with particle size ranging from small to large corresponding to a gradient of blue to yellow (blue represents smaller particles, cyan and green represent intermediate particles, and yellow represents larger particles). Users can intuitively distinguish the distribution of particles of different sizes in the image and quickly determine the uniformity of particle size distribution. At the same time, selected particles are marked with a bold orange outline to highlight the target particles and facilitate focused observation.

[0133] like Figure 5 As shown, the system provides three data display formats: the image display area displays color-coded particle images in real time, annotates the diameter information of the particles, and avoids text overlap through intelligent layout algorithms; the table display area presents parameters such as the ID, caliper diameter, and equivalent diameter of each particle in a list format, supporting sorting and filtering; and the statistics area displays key statistical data such as the total number of particles, D10 / D50 / D90 feature values, average value, and standard deviation, comprehensively presenting the detection results.

[0134] The system supports bidirectional interactive particle selection: when a user clicks on a particle in an image, the system uses a coordinate mapping algorithm to locate the corresponding particle ID, and the parameter row of the corresponding particle in the table is automatically highlighted and scrolled into the visible area; when the user selects a particle row in the table, the corresponding particle in the image is automatically marked with a bold orange outline, achieving synchronized positioning between the image and the table. Furthermore, the system supports batch processing of multiple images and paginated browsing, allowing users to quickly switch between different sample images, compare and analyze detection results, and improve analysis efficiency.

[0135] See Figure 2 The second aspect of this application proposes an artificial intelligence-based metal powder particle size statistics system, including,

[0136] The image acquisition module is configured to acquire preprocessed sample images and obtain high-resolution images containing scale calibration marks;

[0137] The image preprocessing module is configured to preprocess a high-resolution image containing scale calibration marks to obtain a preprocessed grayscale image;

[0138] The particle detection and segmentation module is configured to use the HDP-DSL algorithm with a dual-branch architecture to identify and segment the preprocessed grayscale image to obtain a full-scale high-precision particle segmentation mask.

[0139] The parameter calculation module is configured to calculate the physical feature parameters of a single particle based on a full-scale high-precision particle segmentation mask.

[0140] The statistical analysis module is configured to perform statistical analysis on the physical characteristic parameters of single particles, generate a particle size distribution histogram, and calculate key statistical characteristic values ​​of D10, D50, and D90.

[0141] The visualization and interaction module is configured to visualize the detection results through color coding and interactive design.

[0142] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for statistical analysis of metal powder particle size based on artificial intelligence, characterized in that, Includes the following steps: Step 1: Sample preprocessing and image acquisition to obtain high-resolution images containing scale calibration marks; Step 2: Preprocess the high-resolution image containing scale calibration marks to obtain a preprocessed grayscale image; Step 3: The HDP-DSL algorithm with a dual-branch architecture is used to identify and segment the preprocessed grayscale image to obtain a full-scale high-precision particle segmentation mask. Step 4: Calculate the physical feature parameters of a single particle based on the full-scale high-precision particle segmentation mask; Step 5: Perform statistical analysis on the single-particle physical characteristic parameters output in Step 4, generate a particle size distribution histogram, and calculate key statistical characteristic values ​​such as D10, D50, and D90. Step 6: Visualize the detection results through color coding and interactive design.

2. The method for statistical analysis of metal powder particle size based on artificial intelligence according to claim 1, characterized in that, Step 1 specifically includes: Anhydrous ethanol or acetone is used as a solvent, and the sample is ultrasonically cleaned for 5-10 minutes with a power of 300-500W. After cleaning, the sample is placed in a vacuum drying oven and dried for 2-4 hours at 40-60℃ and a vacuum degree of <10Pa to ensure that the solvent is completely evaporated and there is no residue. Take 1-5 mg of dried metal powder and disperse it evenly on an aluminum or copper conductive adhesive sample stage. If the powder particle size is less than 1 μm, it can be dispersed in anhydrous ethanol to prepare a suspension. Use a micropipette to take 1-2 μL of the suspension and drop it onto the conductive adhesive on the sample stage and vacuum dry it. The powder samples were photographed and observed using a scanning electron microscope. The overall dispersion of the samples was observed at a low magnification of 50-200x, avoiding areas with severe agglomeration or sparse distribution. The magnification was then gradually increased to 500-5000x, and 10-20 representative fields of view were selected for photographing. All photographs were labeled with scale bars. The image resolution was set to 1024×1024 or 2048×2048 pixels and saved in TIFF or PNG format to obtain high-resolution images with scale calibration marks.

3. The method for statistical analysis of metal powder particle size based on artificial intelligence according to claim 1, characterized in that, Step 2 specifically includes: First, the high-resolution image containing scale calibration marks is converted into a grayscale image using an RGB channel weighted fusion algorithm, and then Gaussian blur is applied to the grayscale image to suppress noise. The Laplacian operator is used to enhance the edges of the denoised image. The mathematical expression for edge enhancement is as follows: The edge-enhanced image is combined with the original image through weighted fusion to obtain a preprocessed grayscale image.

4. The method for statistical analysis of metal powder particle size based on artificial intelligence according to claim 1, characterized in that, In step 3, the HDP-DSL algorithm consists of two core modules: a large-size particle detection and segmentation mechanism based on geometric features, and a small-size edge particle detection and segmentation algorithm based on an improved U-Net neural network. The specific operations of the large-size particle detection and segmentation mechanism based on geometric features include: The preprocessed grayscale image is binarized using both global and adaptive local thresholding, and the union of the two thresholding results is taken to obtain candidate contours. For candidate contours, a distance transformation map is obtained using the distance transformation formula. The formula for calculating the distance transformation is: The local maximum detection algorithm is used to find peak points in the distance transformation map. The minimum peak detection distance is set to dmin=0.3max(D(x,y)); if the number of valid peaks detected is greater than 1, the contour is determined to be overlapping particles. For the identified overlapping particles, the overlapping region is segmented into independent particle regions by watershed transformation to obtain the segmentation result of the first branch; Small-size and edge particle detection and segmentation based on an improved U-Net neural network: The preprocessed grayscale image is input into the improved U-Net neural network to obtain the segmentation result of the second branch; The segmentation results of the first branch and the segmentation results of the second branch are complementary and corrected to obtain a full-scale high-precision particle segmentation mask.

5. The method for statistical analysis of metal powder particle size based on artificial intelligence according to claim 4, characterized in that, The improved U-Net neural network is a neural network optimized based on the classic U-Net network architecture; The encoder in the classic U-Net network architecture is optimized by replacing the standard convolution of the classic encoder with depthwise separable convolution. After the encoder feature output, a spatial attention module and a channel attention module are added; Dice Loss and binary cross-entropy loss function are used.

6. The method for statistical analysis of metal powder particle size based on artificial intelligence according to claim 1, characterized in that, Step 4 specifically includes: For each segmented effective particle profile, calculate characteristic parameters such as area, equivalent diameter, caliper diameter, and center coordinates. The equivalent diameter is calculated based on the principle of area equivalence, and the formula is: ; Where deq is the equivalent diameter and A is the particle area. The diameter of the caliper is obtained using the minimum circumscribed rectangle algorithm; the center coordinates are obtained using the minimum circumscribed circle algorithm. Calibration is performed using scale calibration markers in the image, converting pixel-level parameters to physical units.

7. The method for statistical analysis of metal powder particle size based on artificial intelligence according to claim 1, characterized in that, Step 5 specifically includes: Using particle caliper diameter as the core parameter, calculate the minimum, maximum, average, and standard deviation of particle size; The statistical characteristic values ​​of D10, D50, and D90 were calculated using the quantile statistical method. The particle size range is divided into multiple intervals, the number of particles in each interval is counted, and a particle size distribution histogram is generated.

8. A metal powder particle size statistics system based on artificial intelligence, characterized in that, include, The image acquisition module is configured to acquire preprocessed sample images and obtain high-resolution images containing scale calibration marks; The image preprocessing module is configured to preprocess a high-resolution image containing scale calibration marks to obtain a preprocessed grayscale image; The particle detection and segmentation module is configured to use the HDP-DSL algorithm with a dual-branch architecture to identify and segment the preprocessed grayscale image to obtain a full-scale high-precision particle segmentation mask. The parameter calculation module is configured to calculate the physical feature parameters of a single particle based on a full-scale high-precision particle segmentation mask. The statistical analysis module is configured to perform statistical analysis on the physical characteristic parameters of single particles, generate a particle size distribution histogram, and calculate key statistical characteristic values ​​of D10, D50, and D90. The visualization and interaction module is configured to visualize the detection results through color coding and interactive design.

Images